研究库 论文知识库

内容库 / 论文

arXiv:2609.07821 · 安全与风险

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

A*-Thought-V2：通过 LLM 几何动力学实现高效潜在推理

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

类型：arxiv
标识：2609.07821
链接：https://arxiv.org/abs/2609.07821
主分类：risk
形态：method
TLDR：Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local
待LLM分类：否
标题中文：A*-Thought-V2：通过 LLM 几何动力学实现高效潜在推理
TLDR中文：思维链（CoT）提升了大型语言模型（LLM）的推理能力，但带来大量计算与上下文开销。现有方法要么通过硬剪枝丢失中间信息，要么缺乏连续压缩的原则性准则。我们提出 A*-Thought-V2，一个由 LLM 几何动力学引导的框架，将 CoT 建模为隐状态轨迹，并用显式-隐式交替的潜在架构替代硬删除。在将问题、步骤与解的表示投影至三维 PCA 空间后，它度量每个局部……
来源文件：
/inbox/tom/_candidates/2026-09-09-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记精读与批判 · From 128K to 4M: Efficient Training of Ultra-Long Context LLMs

笔记 LLM-Wiki / RAGSearch / GradRAG · 新型 RAG 范式综合 · 2026-07-28

笔记 Sebastian Raschka (Ahead of AI) · RSS 摘要

仓库 Dusto-AI/llm-situational-awareness-paper-list

笔记 Cool Papers · cs.CL (papers.cool) · RSS 摘要

仓库 HanaMekonen/llm-adaptation-mrl-systematic-review

arXiv

2609.07821

TLDR

思维链（CoT）提升了大型语言模型（LLM）的推理能力，但带来大量计算与上下文开销。现有方法要么通过硬剪枝丢失中间信息，要么缺乏连续压缩的原则性准则。我们提出 A*-Thought-V2，一个由 LLM 几何动力学引导的框架，将 CoT 建模为隐状态轨迹，并用显式-隐式交替的潜在架构替代硬删除。在将问题、步骤与解的表示投影至三维 PCA 空间后，它度量每个局部……Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local

arXiv ↗ 中文全文